| 1 |
// SPDX-FileCopyrightText: Copyright © DUNE Project contributors, see file LICENSE.md in module root |
| 2 |
// SPDX-License-Identifier: LicenseRef-GPL-2.0-only-with-DUNE-exception |
| 3 |
// -*- tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 2 -*- |
| 4 |
// vi: set et ts=4 sw=2 sts=2: |
| 5 |
#ifndef DUNE_AMG_AMG_HH |
| 6 |
#define DUNE_AMG_AMG_HH |
| 7 |
|
| 8 |
#include <memory> |
| 9 |
#include <sstream> |
| 10 |
#include <dune/common/exceptions.hh> |
| 11 |
#include <dune/istl/paamg/smoother.hh> |
| 12 |
#include <dune/istl/paamg/transfer.hh> |
| 13 |
#include <dune/istl/paamg/matrixhierarchy.hh> |
| 14 |
#include <dune/istl/solvers.hh> |
| 15 |
#include <dune/istl/scalarproducts.hh> |
| 16 |
#include <dune/istl/superlu.hh> |
| 17 |
#include <dune/istl/umfpack.hh> |
| 18 |
#include <dune/istl/solvertype.hh> |
| 19 |
#include <dune/istl/solverregistry.hh> |
| 20 |
#include <dune/common/typetraits.hh> |
| 21 |
#include <dune/common/exceptions.hh> |
| 22 |
#include <dune/common/scalarvectorview.hh> |
| 23 |
#include <dune/common/scalarmatrixview.hh> |
| 24 |
#include <dune/common/parametertree.hh> |
| 25 |
|
| 26 |
namespace Dune |
| 27 |
{ |
| 28 |
namespace Amg |
| 29 |
{ |
| 30 |
/** |
| 31 |
* @defgroup ISTL_PAAMG Parallel Algebraic Multigrid |
| 32 |
* @ingroup ISTL_Prec |
| 33 |
* @brief A Parallel Algebraic Multigrid based on Agglomeration. |
| 34 |
*/ |
| 35 |
|
| 36 |
/** |
| 37 |
* @addtogroup ISTL_PAAMG |
| 38 |
* |
| 39 |
* @{ |
| 40 |
*/ |
| 41 |
|
| 42 |
/** @file |
| 43 |
* @author Markus Blatt |
| 44 |
* @brief The AMG preconditioner. |
| 45 |
*/ |
| 46 |
|
| 47 |
template<class M, class X, class S, class P, class K, class A> |
| 48 |
class KAMG; |
| 49 |
|
| 50 |
template<class T> |
| 51 |
class KAmgTwoGrid; |
| 52 |
|
| 53 |
/** |
| 54 |
* @brief Parallel algebraic multigrid based on agglomeration. |
| 55 |
* |
| 56 |
* \tparam M The LinearOperator type which represents the matrix |
| 57 |
* \tparam X The vector type |
| 58 |
* \tparam S The smoother type |
| 59 |
* \tparam A An allocator for X |
| 60 |
* |
| 61 |
* \todo drop the smoother template parameter and replace with dynamic construction |
| 62 |
*/ |
| 63 |
template<class M, class X, class S, class PI=SequentialInformation, |
| 64 |
class A=std::allocator<X> > |
| 65 |
class AMG : public Preconditioner<X,X> |
| 66 |
{ |
| 67 |
template<class M1, class X1, class S1, class P1, class K1, class A1> |
| 68 |
friend class KAMG; |
| 69 |
|
| 70 |
friend class KAmgTwoGrid<AMG>; |
| 71 |
|
| 72 |
public: |
| 73 |
/** @brief The matrix operator type. */ |
| 74 |
typedef M Operator; |
| 75 |
/** |
| 76 |
* @brief The type of the parallel information. |
| 77 |
* Either OwnerOverlapCommunication or another type |
| 78 |
* describing the parallel data distribution and |
| 79 |
* providing communication methods. |
| 80 |
*/ |
| 81 |
typedef PI ParallelInformation; |
| 82 |
/** @brief The operator hierarchy type. */ |
| 83 |
typedef MatrixHierarchy<M, ParallelInformation, A> OperatorHierarchy; |
| 84 |
/** @brief The parallal data distribution hierarchy type. */ |
| 85 |
typedef typename OperatorHierarchy::ParallelInformationHierarchy ParallelInformationHierarchy; |
| 86 |
|
| 87 |
/** @brief The domain type. */ |
| 88 |
typedef X Domain; |
| 89 |
/** @brief The range type. */ |
| 90 |
typedef X Range; |
| 91 |
/** @brief the type of the coarse solver. */ |
| 92 |
typedef InverseOperator<X,X> CoarseSolver; |
| 93 |
/** |
| 94 |
* @brief The type of the smoother. |
| 95 |
* |
| 96 |
* One of the preconditioners implementing the Preconditioner interface. |
| 97 |
* Note that the smoother has to fit the ParallelInformation.*/ |
| 98 |
typedef S Smoother; |
| 99 |
|
| 100 |
/** @brief The argument type for the construction of the smoother. */ |
| 101 |
typedef typename SmootherTraits<Smoother>::Arguments SmootherArgs; |
| 102 |
|
| 103 |
/** |
| 104 |
* @brief Construct a new amg with a specific coarse solver. |
| 105 |
* @param matrices The already set-up matrix hierarchy. |
| 106 |
* @param coarseSolver The set up solver to use on the coarse |
| 107 |
* grid, must match the coarse matrix in the matrix hierarchy. |
| 108 |
* @param smootherArgs The arguments needed for thesmoother to use |
| 109 |
* for pre and post smoothing. |
| 110 |
* @param parms The parameters for the AMG. |
| 111 |
*/ |
| 112 |
AMG(OperatorHierarchy& matrices, CoarseSolver& coarseSolver, |
| 113 |
const SmootherArgs& smootherArgs, const Parameters& parms); |
| 114 |
|
| 115 |
/** |
| 116 |
* @brief Construct an AMG with an inexact coarse solver based on the smoother. |
| 117 |
* |
| 118 |
* As coarse solver a preconditioned CG method with the smoother as preconditioner |
| 119 |
* will be used. The matrix hierarchy is built automatically. |
| 120 |
* @param fineOperator The operator on the fine level. |
| 121 |
* @param criterion The criterion describing the coarsening strategy. E. g. SymmetricCriterion |
| 122 |
* or UnsymmetricCriterion, and providing the parameters. |
| 123 |
* @param smootherArgs The arguments for constructing the smoothers. |
| 124 |
* @param pinfo The information about the parallel distribution of the data. |
| 125 |
*/ |
| 126 |
template<class C> |
| 127 |
AMG(const Operator& fineOperator, const C& criterion, |
| 128 |
const SmootherArgs& smootherArgs=SmootherArgs(), |
| 129 |
const ParallelInformation& pinfo=ParallelInformation()); |
| 130 |
|
| 131 |
/*! |
| 132 |
\brief Constructor an AMG via ParameterTree. |
| 133 |
|
| 134 |
\param fineOperator The operator on the fine level. |
| 135 |
\param configuration ParameterTree containing AMG parameters. |
| 136 |
\param pinfo Optionally, specify ParallelInformation |
| 137 |
|
| 138 |
ParameterTree Key | Meaning |
| 139 |
--------------------------|------------ |
| 140 |
smootherIterations | The number of iterations to perform. |
| 141 |
smootherRelaxation | The relaxation factor |
| 142 |
maxLevel | Maximum number of levels allowed in the hierarchy. |
| 143 |
coarsenTarget | Maximum number of unknowns on the coarsest level. |
| 144 |
minCoarseningRate | Coarsening will stop if the rate is below this threshold. |
| 145 |
accumulationMode | If and how data is agglomerated on coarser level to |
| 146 |
| fewer processors. ("atOnce": do agglomeration once and |
| 147 |
| to one process; "successive": Multiple agglomerations to |
| 148 |
| fewer processes until all data is on one process; |
| 149 |
| "none": Do no agglomeration at all and solve coarse level |
| 150 |
| iteratively). |
| 151 |
prolongationDampingFactor | Damping factor for the prolongation. |
| 152 |
alpha | Scaling value for marking connections as strong. |
| 153 |
beta | Threshold for marking nodes as isolated. |
| 154 |
additive | Whether to use additive multigrid. |
| 155 |
gamma | 1 for V-cycle, 2 for W-cycle. |
| 156 |
preSteps | Number of presmoothing steps. |
| 157 |
postSteps | Number of postsmoothing steps. |
| 158 |
verbosity | Output verbosity. default=2. |
| 159 |
criterionSymmetric | If true use SymmetricCriterion (default), else UnSymmetricCriterion |
| 160 |
strengthMeasure | What conversion to use to convert a matrix block |
| 161 |
| to a scalar when determining strength of connection: |
| 162 |
| diagonal (use a diagonal of row diagonalRowIndex, class Diagonal, default); |
| 163 |
| rowSum (rowSum norm), frobenius (Frobenius norm); |
| 164 |
| one (use always one and neglect the actual entries). |
| 165 |
diagonalRowIndex | The index to use for the diagonal strength (default 0) |
| 166 |
| if this is i and strengthMeasure is "diagonal", then |
| 167 |
| block[i][i] will be used when determining strength of |
| 168 |
| connection. |
| 169 |
defaultAggregationSizeMode | Whether to set default values depending on isotropy of |
| 170 |
| problem uses parameters "defaultAggregationDimension" and |
| 171 |
| "maxAggregateDistance" (isotropic: For and isotropic problem; |
| 172 |
| anisotropic: for an anisotropic problem). |
| 173 |
defaultAggregationDimension | Dimension of the problem (used for setting default aggregate size). |
| 174 |
maxAggregateDistance | Maximum distance in an aggregte (in term of minimum edges needed to travel. |
| 175 |
| one vertex to another within the aggregate). |
| 176 |
minAggregateSize | Minimum number of vertices an aggregate should consist of. |
| 177 |
maxAggregateSize | Maximum number of vertices an aggregate should consist of. |
| 178 |
|
| 179 |
See \ref ISTL_Factory for the ParameterTree layout and examples. |
| 180 |
*/ |
| 181 |
AMG(std::shared_ptr<const Operator> fineOperator, const ParameterTree& configuration, const ParallelInformation& pinfo=ParallelInformation()); |
| 182 |
|
| 183 |
/** |
| 184 |
* @brief Copy constructor. |
| 185 |
*/ |
| 186 |
AMG(const AMG& amg); |
| 187 |
|
| 188 |
/** \copydoc Preconditioner::pre */ |
| 189 |
void pre(Domain& x, Range& b); |
| 190 |
|
| 191 |
/** \copydoc Preconditioner::apply */ |
| 192 |
void apply(Domain& v, const Range& d); |
| 193 |
|
| 194 |
//! Category of the preconditioner (see SolverCategory::Category) |
| 195 |
virtual SolverCategory::Category category() const |
| 196 |
{ |
| 197 |
return category_; |
| 198 |
} |
| 199 |
|
| 200 |
/** \copydoc Preconditioner::post */ |
| 201 |
void post(Domain& x); |
| 202 |
|
| 203 |
/** |
| 204 |
* @brief Get the aggregate number of each unknown on the coarsest level. |
| 205 |
* @param cont The random access container to store the numbers in. |
| 206 |
*/ |
| 207 |
template<class A1> |
| 208 |
void getCoarsestAggregateNumbers(std::vector<std::size_t,A1>& cont); |
| 209 |
|
| 210 |
std::size_t levels(); |
| 211 |
|
| 212 |
std::size_t maxlevels(); |
| 213 |
|
| 214 |
/** |
| 215 |
* @brief Recalculate the matrix hierarchy. |
| 216 |
* |
| 217 |
* It is assumed that the coarsening for the changed fine level |
| 218 |
* matrix would yield the same aggregates. In this case it suffices |
| 219 |
* to recalculate all the Galerkin products for the matrices of the |
| 220 |
* coarser levels. |
| 221 |
*/ |
| 222 |
void recalculateHierarchy() |
| 223 |
{ |
| 224 |
matrices_->recalculateGalerkin(NegateSet<typename PI::OwnerSet>()); |
| 225 |
} |
| 226 |
|
| 227 |
/** |
| 228 |
* @brief Check whether the coarse solver used is a direct solver. |
| 229 |
* @return True if the coarse level solver is a direct solver. |
| 230 |
*/ |
| 231 |
bool usesDirectCoarseLevelSolver() const; |
| 232 |
|
| 233 |
private: |
| 234 |
/* |
| 235 |
* @brief Helper function to create hierarchies with parameter tree. |
| 236 |
* |
| 237 |
* Will create the coarsen criterion with the norm and create the |
| 238 |
* Hierarchies |
| 239 |
* \tparam Norm Type of the norm to use. |
| 240 |
*/ |
| 241 |
template<class Norm> |
| 242 |
void createCriterionAndHierarchies(std::shared_ptr<const Operator> matrixptr, |
| 243 |
const PI& pinfo, const Norm&, |
| 244 |
const ParameterTree& configuration, |
| 245 |
std::true_type compiles = std::true_type()); |
| 246 |
template<class Norm> |
| 247 |
void createCriterionAndHierarchies(std::shared_ptr<const Operator> matrixptr, |
| 248 |
const PI& pinfo, const Norm&, |
| 249 |
const ParameterTree& configuration, |
| 250 |
std::false_type); |
| 251 |
/** |
| 252 |
* @brief Helper function to create hierarchies with settings from parameter tree. |
| 253 |
* @param criterion Coarsen criterion to configure and use |
| 254 |
*/ |
| 255 |
template<class C> |
| 256 |
void createHierarchies(C& criterion, std::shared_ptr<const Operator> matrixptr, |
| 257 |
const PI& pinfo, const ParameterTree& configuration); |
| 258 |
/** |
| 259 |
* @brief Create matrix and smoother hierarchies. |
| 260 |
* @param criterion The coarsening criterion. |
| 261 |
* @param matrix The fine level matrix operator. |
| 262 |
* @param pinfo The fine level parallel information. |
| 263 |
*/ |
| 264 |
template<class C> |
| 265 |
void createHierarchies(C& criterion, |
| 266 |
const std::shared_ptr<const Operator>& matrixptr, |
| 267 |
const PI& pinfo); |
| 268 |
/** |
| 269 |
* @brief A struct that holds the context of the current level. |
| 270 |
* |
| 271 |
* These are the iterators to the smoother, matrix, parallel information, |
| 272 |
* and so on needed for the computations on the current level. |
| 273 |
*/ |
| 274 |
struct LevelContext |
| 275 |
{ |
| 276 |
typedef Smoother SmootherType; |
| 277 |
/** |
| 278 |
* @brief The iterator over the smoothers. |
| 279 |
*/ |
| 280 |
typename Hierarchy<Smoother,A>::Iterator smoother; |
| 281 |
/** |
| 282 |
* @brief The iterator over the matrices. |
| 283 |
*/ |
| 284 |
typename OperatorHierarchy::ParallelMatrixHierarchy::ConstIterator matrix; |
| 285 |
/** |
| 286 |
* @brief The iterator over the parallel information. |
| 287 |
*/ |
| 288 |
typename ParallelInformationHierarchy::Iterator pinfo; |
| 289 |
/** |
| 290 |
* @brief The iterator over the redistribution information. |
| 291 |
*/ |
| 292 |
typename OperatorHierarchy::RedistributeInfoList::const_iterator redist; |
| 293 |
/** |
| 294 |
* @brief The iterator over the aggregates maps. |
| 295 |
*/ |
| 296 |
typename OperatorHierarchy::AggregatesMapList::const_iterator aggregates; |
| 297 |
/** |
| 298 |
* @brief The iterator over the left hand side. |
| 299 |
*/ |
| 300 |
typename Hierarchy<Domain,A>::Iterator lhs; |
| 301 |
/** |
| 302 |
* @brief The iterator over the updates. |
| 303 |
*/ |
| 304 |
typename Hierarchy<Domain,A>::Iterator update; |
| 305 |
/** |
| 306 |
* @brief The iterator over the right hand sided. |
| 307 |
*/ |
| 308 |
typename Hierarchy<Range,A>::Iterator rhs; |
| 309 |
/** |
| 310 |
* @brief The level index. |
| 311 |
*/ |
| 312 |
std::size_t level; |
| 313 |
}; |
| 314 |
|
| 315 |
|
| 316 |
/** |
| 317 |
* @brief Multigrid cycle on a level. |
| 318 |
* @param levelContext the iterators of the current level. |
| 319 |
*/ |
| 320 |
void mgc(LevelContext& levelContext); |
| 321 |
|
| 322 |
void additiveMgc(); |
| 323 |
|
| 324 |
/** |
| 325 |
* @brief Move the iterators to the finer level |
| 326 |
* @param levelContext the iterators of the current level. |
| 327 |
* @param processedFineLevel Whether the process computed on |
| 328 |
* fine level or not. |
| 329 |
*/ |
| 330 |
void moveToFineLevel(LevelContext& levelContext,bool processedFineLevel); |
| 331 |
|
| 332 |
/** |
| 333 |
* @brief Move the iterators to the coarser level. |
| 334 |
* @param levelContext the iterators of the current level |
| 335 |
*/ |
| 336 |
bool moveToCoarseLevel(LevelContext& levelContext); |
| 337 |
|
| 338 |
/** |
| 339 |
* @brief Initialize iterators over levels with fine level. |
| 340 |
* @param levelContext the iterators of the current level |
| 341 |
*/ |
| 342 |
void initIteratorsWithFineLevel(LevelContext& levelContext); |
| 343 |
|
| 344 |
/** @brief The matrix we solve. */ |
| 345 |
std::shared_ptr<OperatorHierarchy> matrices_; |
| 346 |
/** @brief The arguments to construct the smoother */ |
| 347 |
SmootherArgs smootherArgs_; |
| 348 |
/** @brief The hierarchy of the smoothers. */ |
| 349 |
std::shared_ptr<Hierarchy<Smoother,A> > smoothers_; |
| 350 |
/** @brief The solver of the coarsest level. */ |
| 351 |
std::shared_ptr<CoarseSolver> solver_; |
| 352 |
/** @brief The right hand side of our problem. */ |
| 353 |
std::shared_ptr<Hierarchy<Range,A>> rhs_; |
| 354 |
/** @brief The left approximate solution of our problem. */ |
| 355 |
std::shared_ptr<Hierarchy<Domain,A>> lhs_; |
| 356 |
/** @brief The total update for the outer solver. */ |
| 357 |
std::shared_ptr<Hierarchy<Domain,A>> update_; |
| 358 |
/** @brief The type of the scalar product for the coarse solver. */ |
| 359 |
using ScalarProduct = Dune::ScalarProduct<X>; |
| 360 |
/** @brief Scalar product on the coarse level. */ |
| 361 |
std::shared_ptr<ScalarProduct> scalarProduct_; |
| 362 |
/** @brief Gamma, 1 for V-cycle and 2 for W-cycle. */ |
| 363 |
std::size_t gamma_; |
| 364 |
/** @brief The number of pre and postsmoothing steps. */ |
| 365 |
std::size_t preSteps_; |
| 366 |
/** @brief The number of postsmoothing steps. */ |
| 367 |
std::size_t postSteps_; |
| 368 |
bool buildHierarchy_; |
| 369 |
bool additive; |
| 370 |
bool coarsesolverconverged; |
| 371 |
std::shared_ptr<Smoother> coarseSmoother_; |
| 372 |
/** @brief The solver category. */ |
| 373 |
SolverCategory::Category category_; |
| 374 |
/** @brief The verbosity level. */ |
| 375 |
std::size_t verbosity_; |
| 376 |
|
| 377 |
struct ToLower |
| 378 |
{ |
| 379 |
std::string operator()(const std::string& str) |
| 380 |
{ |
| 381 |
std::stringstream retval; |
| 382 |
std::ostream_iterator<char> out(retval); |
| 383 |
std::transform(str.begin(), str.end(), out, |
| 384 |
[](char c){ |
| 385 |
return std::tolower(c, std::locale::classic()); |
| 386 |
}); |
| 387 |
return retval.str(); |
| 388 |
} |
| 389 |
}; |
| 390 |
}; |
| 391 |
|
| 392 |
template<class M, class X, class S, class PI, class A> |
| 393 |
inline AMG<M,X,S,PI,A>::AMG(const AMG& amg) |
| 394 |
: matrices_(amg.matrices_), smootherArgs_(amg.smootherArgs_), |
| 395 |
smoothers_(amg.smoothers_), solver_(amg.solver_), |
| 396 |
rhs_(), lhs_(), update_(), |
| 397 |
scalarProduct_(amg.scalarProduct_), gamma_(amg.gamma_), |
| 398 |
preSteps_(amg.preSteps_), postSteps_(amg.postSteps_), |
| 399 |
buildHierarchy_(amg.buildHierarchy_), |
| 400 |
additive(amg.additive), coarsesolverconverged(amg.coarsesolverconverged), |
| 401 |
coarseSmoother_(amg.coarseSmoother_), |
| 402 |
category_(amg.category_), |
| 403 |
verbosity_(amg.verbosity_) |
| 404 |
{} |
| 405 |
|
| 406 |
template<class M, class X, class S, class PI, class A> |
| 407 |
AMG<M,X,S,PI,A>::AMG(OperatorHierarchy& matrices, CoarseSolver& coarseSolver, |
| 408 |
const SmootherArgs& smootherArgs, |
| 409 |
const Parameters& parms) |
| 410 |
: matrices_(stackobject_to_shared_ptr(matrices)), smootherArgs_(smootherArgs), |
| 411 |
smoothers_(new Hierarchy<Smoother,A>), solver_(&coarseSolver), |
| 412 |
rhs_(), lhs_(), update_(), scalarProduct_(0), |
| 413 |
gamma_(parms.getGamma()), preSteps_(parms.getNoPreSmoothSteps()), |
| 414 |
postSteps_(parms.getNoPostSmoothSteps()), buildHierarchy_(false), |
| 415 |
additive(parms.getAdditive()), coarsesolverconverged(true), |
| 416 |
coarseSmoother_(), |
| 417 |
// #warning should category be retrieved from matrices? |
| 418 |
category_(SolverCategory::category(*smoothers_->coarsest())), |
| 419 |
verbosity_(parms.debugLevel()) |
| 420 |
{ |
| 421 |
assert(matrices_->isBuilt()); |
| 422 |
|
| 423 |
// build the necessary smoother hierarchies |
| 424 |
matrices_->coarsenSmoother(*smoothers_, smootherArgs_); |
| 425 |
} |
| 426 |
|
| 427 |
template<class M, class X, class S, class PI, class A> |
| 428 |
template<class C> |
| 429 |
AMG<M,X,S,PI,A>::AMG(const Operator& matrix, |
| 430 |
const C& criterion, |
| 431 |
const SmootherArgs& smootherArgs, |
| 432 |
const PI& pinfo) |
| 433 |
: smootherArgs_(smootherArgs), |
| 434 |
smoothers_(new Hierarchy<Smoother,A>), solver_(), |
| 435 |
rhs_(), lhs_(), update_(), scalarProduct_(), |
| 436 |
gamma_(criterion.getGamma()), preSteps_(criterion.getNoPreSmoothSteps()), |
| 437 |
postSteps_(criterion.getNoPostSmoothSteps()), buildHierarchy_(true), |
| 438 |
additive(criterion.getAdditive()), coarsesolverconverged(true), |
| 439 |
coarseSmoother_(), |
| 440 |
category_(SolverCategory::category(pinfo)), |
| 441 |
verbosity_(criterion.debugLevel()) |
| 442 |
{ |
| 443 |
if(SolverCategory::category(matrix) != SolverCategory::category(pinfo)) |
| 444 |
DUNE_THROW(InvalidSolverCategory, "Matrix and Communication must have the same SolverCategory!"); |
| 445 |
// TODO: reestablish compile time checks. |
| 446 |
//static_assert(static_cast<int>(PI::category)==static_cast<int>(S::category), |
| 447 |
// "Matrix and Solver must match in terms of category!"); |
| 448 |
auto matrixptr = stackobject_to_shared_ptr(matrix); |
| 449 |
createHierarchies(criterion, matrixptr, pinfo); |
| 450 |
} |
| 451 |
|
| 452 |
template<class M, class X, class S, class PI, class A> |
| 453 |
AMG<M,X,S,PI,A>::AMG(std::shared_ptr<const Operator> matrixptr, |
| 454 |
const ParameterTree& configuration, |
| 455 |
const ParallelInformation& pinfo) : |
| 456 |
smoothers_(new Hierarchy<Smoother,A>), |
| 457 |
solver_(), rhs_(), lhs_(), update_(), scalarProduct_(), buildHierarchy_(true), |
| 458 |
coarsesolverconverged(true), coarseSmoother_(), |
| 459 |
category_(SolverCategory::category(pinfo)) |
| 460 |
{ |
| 461 |
|
| 462 |
if (configuration.hasKey ("smootherIterations")) |
| 463 |
smootherArgs_.iterations = configuration.get<int>("smootherIterations"); |
| 464 |
|
| 465 |
if (configuration.hasKey ("smootherRelaxation")) |
| 466 |
smootherArgs_.relaxationFactor = configuration.get<typename SmootherArgs::RelaxationFactor>("smootherRelaxation"); |
| 467 |
|
| 468 |
auto normName = ToLower()(configuration.get("strengthMeasure", "diagonal")); |
| 469 |
auto index = configuration.get<int>("diagonalRowIndex", 0); |
| 470 |
|
| 471 |
if ( normName == "diagonal") |
| 472 |
{ |
| 473 |
using field_type = typename M::field_type; |
| 474 |
using real_type = typename FieldTraits<field_type>::real_type; |
| 475 |
std::is_convertible<field_type, real_type> compiles; |
| 476 |
|
| 477 |
switch (index) |
| 478 |
{ |
| 479 |
case 0: |
| 480 |
createCriterionAndHierarchies(matrixptr, pinfo, Diagonal<0>(), configuration, compiles); |
| 481 |
break; |
| 482 |
case 1: |
| 483 |
createCriterionAndHierarchies(matrixptr, pinfo, Diagonal<1>(), configuration, compiles); |
| 484 |
break; |
| 485 |
case 2: |
| 486 |
createCriterionAndHierarchies(matrixptr, pinfo, Diagonal<2>(), configuration, compiles); |
| 487 |
break; |
| 488 |
case 3: |
| 489 |
createCriterionAndHierarchies(matrixptr, pinfo, Diagonal<3>(), configuration, compiles); |
| 490 |
break; |
| 491 |
case 4: |
| 492 |
createCriterionAndHierarchies(matrixptr, pinfo, Diagonal<4>(), configuration, compiles); |
| 493 |
break; |
| 494 |
default: |
| 495 |
DUNE_THROW(InvalidStateException, "Currently strengthIndex>4 is not supported."); |
| 496 |
} |
| 497 |
} |
| 498 |
else if (normName == "rowsum") |
| 499 |
createCriterionAndHierarchies(matrixptr, pinfo, RowSum(), configuration); |
| 500 |
else if (normName == "frobenius") |
| 501 |
createCriterionAndHierarchies(matrixptr, pinfo, FrobeniusNorm(), configuration); |
| 502 |
else if (normName == "one") |
| 503 |
createCriterionAndHierarchies(matrixptr, pinfo, AlwaysOneNorm(), configuration); |
| 504 |
else |
| 505 |
DUNE_THROW(Dune::NotImplemented, "Wrong config file: strengthMeasure "<<normName<<" is not supported by AMG"); |
| 506 |
} |
| 507 |
|
| 508 |
template<class M, class X, class S, class PI, class A> |
| 509 |
template<class Norm> |
| 510 |
void AMG<M,X,S,PI,A>::createCriterionAndHierarchies(std::shared_ptr<const Operator> matrixptr, const PI& pinfo, const Norm&, const ParameterTree& configuration, std::false_type) |
| 511 |
{ |
| 512 |
DUNE_THROW(InvalidStateException, "Strength of connection measure does not support this type (" |
| 513 |
<< className<typename M::field_type>() << ") as it is lacking a conversion to" |
| 514 |
<< className<typename FieldTraits<typename M::field_type>::real_type>() << "."); |
| 515 |
} |
| 516 |
|
| 517 |
template<class M, class X, class S, class PI, class A> |
| 518 |
template<class Norm> |
| 519 |
void AMG<M,X,S,PI,A>::createCriterionAndHierarchies(std::shared_ptr<const Operator> matrixptr, const PI& pinfo, const Norm&, const ParameterTree& configuration, std::true_type) |
| 520 |
{ |
| 521 |
if (configuration.get<bool>("criterionSymmetric", true)) |
| 522 |
{ |
| 523 |
using Criterion = Dune::Amg::CoarsenCriterion< |
| 524 |
Dune::Amg::SymmetricCriterion<typename M::matrix_type,Norm> >; |
| 525 |
Criterion criterion; |
| 526 |
createHierarchies(criterion, matrixptr, pinfo, configuration); |
| 527 |
} |
| 528 |
else |
| 529 |
{ |
| 530 |
using Criterion = Dune::Amg::CoarsenCriterion< |
| 531 |
Dune::Amg::UnSymmetricCriterion<typename M::matrix_type,Norm> >; |
| 532 |
Criterion criterion; |
| 533 |
createHierarchies(criterion, matrixptr, pinfo, configuration); |
| 534 |
} |
| 535 |
} |
| 536 |
|
| 537 |
template<class M, class X, class S, class PI, class A> |
| 538 |
template<class C> |
| 539 |
void AMG<M,X,S,PI,A>::createHierarchies(C& criterion, std::shared_ptr<const Operator> matrixptr, const PI& pinfo, const ParameterTree& configuration) |
| 540 |
{ |
| 541 |
if (configuration.hasKey ("maxLevel")) |
| 542 |
criterion.setMaxLevel(configuration.get<int>("maxLevel")); |
| 543 |
|
| 544 |
if (configuration.hasKey ("minCoarseningRate")) |
| 545 |
criterion.setMinCoarsenRate(configuration.get<int>("minCoarseningRate")); |
| 546 |
|
| 547 |
if (configuration.hasKey ("coarsenTarget")) |
| 548 |
criterion.setCoarsenTarget (configuration.get<int>("coarsenTarget")); |
| 549 |
|
| 550 |
if (configuration.hasKey ("accumulationMode")) |
| 551 |
{ |
| 552 |
std::string mode = ToLower()(configuration.get<std::string>("accumulationMode")); |
| 553 |
if ( mode == "none") |
| 554 |
criterion.setAccumulate(AccumulationMode::noAccu); |
| 555 |
else if ( mode == "atonce" ) |
| 556 |
criterion.setAccumulate(AccumulationMode::atOnceAccu); |
| 557 |
else if ( mode == "successive") |
| 558 |
criterion.setCoarsenTarget (AccumulationMode::successiveAccu); |
| 559 |
else |
| 560 |
DUNE_THROW(InvalidSolverFactoryConfiguration, "Parameter accumulationMode does not allow value " |
| 561 |
<< mode <<"."); |
| 562 |
} |
| 563 |
|
| 564 |
if (configuration.hasKey ("prolongationDampingFactor")) |
| 565 |
criterion.setProlongationDampingFactor (configuration.get<double>("prolongationDampingFactor")); |
| 566 |
|
| 567 |
if (configuration.hasKey("defaultAggregationSizeMode")) |
| 568 |
{ |
| 569 |
auto mode = ToLower()(configuration.get<std::string>("defaultAggregationSizeMode")); |
| 570 |
auto dim = configuration.get<std::size_t>("defaultAggregationDimension"); |
| 571 |
std::size_t maxDistance = 2; |
| 572 |
if (configuration.hasKey("MaxAggregateDistance")) |
| 573 |
maxDistance = configuration.get<std::size_t>("maxAggregateDistance"); |
| 574 |
if (mode == "isotropic") |
| 575 |
criterion.setDefaultValuesIsotropic(dim, maxDistance); |
| 576 |
else if(mode == "anisotropic") |
| 577 |
criterion.setDefaultValuesAnisotropic(dim, maxDistance); |
| 578 |
else |
| 579 |
DUNE_THROW(InvalidSolverFactoryConfiguration, "Parameter accumulationMode does not allow value " |
| 580 |
<< mode <<"."); |
| 581 |
} |
| 582 |
|
| 583 |
if (configuration.hasKey("maxAggregateDistance")) |
| 584 |
criterion.setMaxDistance(configuration.get<std::size_t>("maxAggregateDistance")); |
| 585 |
|
| 586 |
if (configuration.hasKey("minAggregateSize")) |
| 587 |
criterion.setMinAggregateSize(configuration.get<std::size_t>("minAggregateSize")); |
| 588 |
|
| 589 |
if (configuration.hasKey("maxAggregateSize")) |
| 590 |
criterion.setMaxAggregateSize(configuration.get<std::size_t>("maxAggregateSize")); |
| 591 |
|
| 592 |
if (configuration.hasKey("maxAggregateConnectivity")) |
| 593 |
criterion.setMaxConnectivity(configuration.get<std::size_t>("maxAggregateConnectivity")); |
| 594 |
|
| 595 |
if (configuration.hasKey ("alpha")) |
| 596 |
criterion.setAlpha (configuration.get<double> ("alpha")); |
| 597 |
|
| 598 |
if (configuration.hasKey ("beta")) |
| 599 |
criterion.setBeta (configuration.get<double> ("beta")); |
| 600 |
|
| 601 |
if (configuration.hasKey ("gamma")) |
| 602 |
criterion.setGamma (configuration.get<std::size_t> ("gamma")); |
| 603 |
gamma_ = criterion.getGamma(); |
| 604 |
|
| 605 |
if (configuration.hasKey ("additive")) |
| 606 |
criterion.setAdditive (configuration.get<bool>("additive")); |
| 607 |
additive = criterion.getAdditive(); |
| 608 |
|
| 609 |
if (configuration.hasKey ("preSteps")) |
| 610 |
criterion.setNoPreSmoothSteps (configuration.get<std::size_t> ("preSteps")); |
| 611 |
preSteps_ = criterion.getNoPreSmoothSteps (); |
| 612 |
|
| 613 |
if (configuration.hasKey ("postSteps")) |
| 614 |
criterion.setNoPostSmoothSteps (configuration.get<std::size_t> ("postSteps")); |
| 615 |
postSteps_ = criterion.getNoPostSmoothSteps (); |
| 616 |
|
| 617 |
verbosity_ = configuration.get("verbosity", 0); |
| 618 |
criterion.setDebugLevel (verbosity_); |
| 619 |
|
| 620 |
createHierarchies(criterion, matrixptr, pinfo); |
| 621 |
} |
| 622 |
|
| 623 |
template <class Matrix, |
| 624 |
class Vector> |
| 625 |
struct DirectSolverSelector |
| 626 |
{ |
| 627 |
typedef typename Matrix :: field_type field_type; |
| 628 |
enum SolverType { umfpack, superlu, none }; |
| 629 |
|
| 630 |
static constexpr SolverType solver = |
| 631 |
#if DISABLE_AMG_DIRECTSOLVER |
| 632 |
none; |
| 633 |
#elif HAVE_SUITESPARSE_UMFPACK |
| 634 |
UMFPackMethodChooser< field_type > :: valid ? umfpack : none ; |
| 635 |
#elif HAVE_SUPERLU |
| 636 |
superlu ; |
| 637 |
#else |
| 638 |
none; |
| 639 |
#endif |
| 640 |
|
| 641 |
template <class M, SolverType> |
| 642 |
struct Solver |
| 643 |
{ |
| 644 |
typedef InverseOperator<Vector,Vector> type; |
| 645 |
static type* create(const M& mat, bool verbose, bool reusevector ) |
| 646 |
{ |
| 647 |
DUNE_THROW(NotImplemented,"DirectSolver not selected"); |
| 648 |
return nullptr; |
| 649 |
} |
| 650 |
static std::string name () { return "None"; } |
| 651 |
}; |
| 652 |
#if HAVE_SUITESPARSE_UMFPACK |
| 653 |
template <class M> |
| 654 |
struct Solver< M, umfpack > |
| 655 |
{ |
| 656 |
typedef UMFPack< M > type; |
| 657 |
static type* create(const M& mat, bool verbose, bool reusevector ) |
| 658 |
{ |
| 659 |
return new type(mat, verbose, reusevector ); |
| 660 |
} |
| 661 |
static std::string name () { return "UMFPack"; } |
| 662 |
}; |
| 663 |
#endif |
| 664 |
#if HAVE_SUPERLU |
| 665 |
template <class M> |
| 666 |
struct Solver< M, superlu > |
| 667 |
{ |
| 668 |
typedef SuperLU< M > type; |
| 669 |
static type* create(const M& mat, bool verbose, bool reusevector ) |
| 670 |
{ |
| 671 |
return new type(mat, verbose, reusevector ); |
| 672 |
} |
| 673 |
static std::string name () { return "SuperLU"; } |
| 674 |
}; |
| 675 |
#endif |
| 676 |
|
| 677 |
// define direct solver type to be used |
| 678 |
typedef Solver< Matrix, solver > SelectedSolver ; |
| 679 |
typedef typename SelectedSolver :: type DirectSolver; |
| 680 |
static constexpr bool isDirectSolver = solver != none; |
| 681 |
static std::string name() { return SelectedSolver :: name (); } |
| 682 |
static DirectSolver* create(const Matrix& mat, bool verbose, bool reusevector ) |
| 683 |
{ |
| 684 |
return SelectedSolver :: create( mat, verbose, reusevector ); |
| 685 |
} |
| 686 |
}; |
| 687 |
|
| 688 |
template<class M, class X, class S, class PI, class A> |
| 689 |
template<class C> |
| 690 |
void AMG<M,X,S,PI,A>::createHierarchies(C& criterion, |
| 691 |
const std::shared_ptr<const Operator>& matrixptr, |
| 692 |
const PI& pinfo) |
| 693 |
{ |
| 694 |
Timer watch; |
| 695 |
matrices_ = std::make_shared<OperatorHierarchy>( |
| 696 |
std::const_pointer_cast<Operator>(matrixptr), |
| 697 |
stackobject_to_shared_ptr(const_cast<PI&>(pinfo))); |
| 698 |
|
| 699 |
matrices_->template build<NegateSet<typename PI::OwnerSet> >(criterion); |
| 700 |
|
| 701 |
// build the necessary smoother hierarchies |
| 702 |
matrices_->coarsenSmoother(*smoothers_, smootherArgs_); |
| 703 |
|
| 704 |
// test whether we should solve on the coarse level. That is the case if we |
| 705 |
// have that level and if there was a redistribution on this level then our |
| 706 |
// communicator has to be valid (size()>0) as the smoother might try to communicate |
| 707 |
// in the constructor. |
| 708 |
if(buildHierarchy_ && matrices_->levels()==matrices_->maxlevels() |
| 709 |
&& ( ! matrices_->redistributeInformation().back().isSetup() || |
| 710 |
matrices_->parallelInformation().coarsest().getRedistributed().communicator().size() ) ) |
| 711 |
{ |
| 712 |
// We have the carsest level. Create the coarse Solver |
| 713 |
SmootherArgs sargs(smootherArgs_); |
| 714 |
sargs.iterations = 1; |
| 715 |
|
| 716 |
typename ConstructionTraits<Smoother>::Arguments cargs; |
| 717 |
cargs.setArgs(sargs); |
| 718 |
if(matrices_->redistributeInformation().back().isSetup()) { |
| 719 |
// Solve on the redistributed partitioning |
| 720 |
cargs.setMatrix(matrices_->matrices().coarsest().getRedistributed().getmat()); |
| 721 |
cargs.setComm(matrices_->parallelInformation().coarsest().getRedistributed()); |
| 722 |
}else{ |
| 723 |
cargs.setMatrix(matrices_->matrices().coarsest()->getmat()); |
| 724 |
cargs.setComm(*matrices_->parallelInformation().coarsest()); |
| 725 |
} |
| 726 |
|
| 727 |
coarseSmoother_ = ConstructionTraits<Smoother>::construct(cargs); |
| 728 |
scalarProduct_ = createScalarProduct<X>(cargs.getComm(),category()); |
| 729 |
|
| 730 |
typedef DirectSolverSelector< typename M::matrix_type, X > SolverSelector; |
| 731 |
|
| 732 |
// Use superlu if we are purely sequential or with only one processor on the coarsest level. |
| 733 |
if( SolverSelector::isDirectSolver && |
| 734 |
(std::is_same<ParallelInformation,SequentialInformation>::value // sequential mode |
| 735 |
|| matrices_->parallelInformation().coarsest()->communicator().size()==1 //parallel mode and only one processor |
| 736 |
|| (matrices_->parallelInformation().coarsest().isRedistributed() |
| 737 |
&& matrices_->parallelInformation().coarsest().getRedistributed().communicator().size()==1 |
| 738 |
&& matrices_->parallelInformation().coarsest().getRedistributed().communicator().size()>0) ) |
| 739 |
) |
| 740 |
{ // redistribute and 1 proc |
| 741 |
if(matrices_->parallelInformation().coarsest().isRedistributed()) |
| 742 |
{ |
| 743 |
if(matrices_->matrices().coarsest().getRedistributed().getmat().N()>0) |
| 744 |
{ |
| 745 |
// We are still participating on this level |
| 746 |
solver_.reset(SolverSelector::create(matrices_->matrices().coarsest().getRedistributed().getmat(), false, false)); |
| 747 |
} |
| 748 |
else |
| 749 |
solver_.reset(); |
| 750 |
} |
| 751 |
else |
| 752 |
{ |
| 753 |
solver_.reset(SolverSelector::create(matrices_->matrices().coarsest()->getmat(), false, false)); |
| 754 |
} |
| 755 |
if(verbosity_>0 && matrices_->parallelInformation().coarsest()->communicator().rank()==0) |
| 756 |
std::cout<< "Using a direct coarse solver (" << SolverSelector::name() << ")" << std::endl; |
| 757 |
} |
| 758 |
else |
| 759 |
{ |
| 760 |
if(matrices_->parallelInformation().coarsest().isRedistributed()) |
| 761 |
{ |
| 762 |
if(matrices_->matrices().coarsest().getRedistributed().getmat().N()>0) |
| 763 |
// We are still participating on this level |
| 764 |
|
| 765 |
// we have to allocate these types using the rebound allocator |
| 766 |
// in order to ensure that we fulfill the alignment requirements |
| 767 |
solver_.reset(new BiCGSTABSolver<X>(const_cast<M&>(matrices_->matrices().coarsest().getRedistributed()), |
| 768 |
*scalarProduct_, |
| 769 |
*coarseSmoother_, 1E-2, 1000, 0)); |
| 770 |
else |
| 771 |
solver_.reset(); |
| 772 |
}else |
| 773 |
{ |
| 774 |
solver_.reset(new BiCGSTABSolver<X>(const_cast<M&>(*matrices_->matrices().coarsest()), |
| 775 |
*scalarProduct_, |
| 776 |
*coarseSmoother_, 1E-2, 1000, 0)); |
| 777 |
// // we have to allocate these types using the rebound allocator |
| 778 |
// // in order to ensure that we fulfill the alignment requirements |
| 779 |
// using Alloc = typename std::allocator_traits<A>::template rebind_alloc<BiCGSTABSolver<X>>; |
| 780 |
// Alloc alloc; |
| 781 |
// auto p = alloc.allocate(1); |
| 782 |
// std::allocator_traits<Alloc>::construct(alloc, p, |
| 783 |
// const_cast<M&>(*matrices_->matrices().coarsest()), |
| 784 |
// *scalarProduct_, |
| 785 |
// *coarseSmoother_, 1E-2, 1000, 0); |
| 786 |
// solver_.reset(p,[](BiCGSTABSolver<X>* p){ |
| 787 |
// Alloc alloc; |
| 788 |
// std::allocator_traits<Alloc>::destroy(alloc, p); |
| 789 |
// alloc.deallocate(p,1); |
| 790 |
// }); |
| 791 |
} |
| 792 |
} |
| 793 |
} |
| 794 |
|
| 795 |
if(verbosity_>0 && matrices_->parallelInformation().finest()->communicator().rank()==0) |
| 796 |
std::cout<<"Building hierarchy of "<<matrices_->maxlevels()<<" levels " |
| 797 |
<<"(including coarse solver) took "<<watch.elapsed()<<" seconds."<<std::endl; |
| 798 |
} |
| 799 |
|
| 800 |
|
| 801 |
template<class M, class X, class S, class PI, class A> |
| 802 |
void AMG<M,X,S,PI,A>::pre(Domain& x, Range& b) |
| 803 |
{ |
| 804 |
// Detect Matrix rows where all offdiagonal entries are |
| 805 |
// zero and set x such that A_dd*x_d=b_d |
| 806 |
// Thus users can be more careless when setting up their linear |
| 807 |
// systems. |
| 808 |
typedef typename M::matrix_type Matrix; |
| 809 |
typedef typename Matrix::ConstRowIterator RowIter; |
| 810 |
typedef typename Matrix::ConstColIterator ColIter; |
| 811 |
typedef typename Matrix::block_type Block; |
| 812 |
Block zero; |
| 813 |
zero=typename Matrix::field_type(); |
| 814 |
|
| 815 |
const Matrix& mat=matrices_->matrices().finest()->getmat(); |
| 816 |
for(RowIter row=mat.begin(); row!=mat.end(); ++row) { |
| 817 |
bool isDirichlet = true; |
| 818 |
bool hasDiagonal = false; |
| 819 |
Block diagonal{}; |
| 820 |
for(ColIter col=row->begin(); col!=row->end(); ++col) { |
| 821 |
if(row.index()==col.index()) { |
| 822 |
diagonal = *col; |
| 823 |
hasDiagonal = true; |
| 824 |
}else{ |
| 825 |
if(*col!=zero) |
| 826 |
isDirichlet = false; |
| 827 |
} |
| 828 |
} |
| 829 |
if(isDirichlet && hasDiagonal) |
| 830 |
{ |
| 831 |
auto&& xEntry = Impl::asVector(x[row.index()]); |
| 832 |
auto&& bEntry = Impl::asVector(b[row.index()]); |
| 833 |
Impl::asMatrix(diagonal).solve(xEntry, bEntry); |
| 834 |
} |
| 835 |
} |
| 836 |
|
| 837 |
if(smoothers_->levels()>0) |
| 838 |
smoothers_->finest()->pre(x,b); |
| 839 |
else |
| 840 |
// No smoother to make x consistent! Do it by hand |
| 841 |
matrices_->parallelInformation().coarsest()->copyOwnerToAll(x,x); |
| 842 |
rhs_ = std::make_shared<Hierarchy<Range,A>>(std::make_shared<Range>(b)); |
| 843 |
lhs_ = std::make_shared<Hierarchy<Domain,A>>(std::make_shared<Domain>(x)); |
| 844 |
update_ = std::make_shared<Hierarchy<Domain,A>>(std::make_shared<Domain>(x)); |
| 845 |
matrices_->coarsenVector(*rhs_); |
| 846 |
matrices_->coarsenVector(*lhs_); |
| 847 |
matrices_->coarsenVector(*update_); |
| 848 |
|
| 849 |
// Preprocess all smoothers |
| 850 |
typedef typename Hierarchy<Smoother,A>::Iterator Iterator; |
| 851 |
typedef typename Hierarchy<Range,A>::Iterator RIterator; |
| 852 |
typedef typename Hierarchy<Domain,A>::Iterator DIterator; |
| 853 |
Iterator coarsest = smoothers_->coarsest(); |
| 854 |
Iterator smoother = smoothers_->finest(); |
| 855 |
RIterator rhs = rhs_->finest(); |
| 856 |
DIterator lhs = lhs_->finest(); |
| 857 |
if(smoothers_->levels()>1) { |
| 858 |
|
| 859 |
assert(lhs_->levels()==rhs_->levels()); |
| 860 |
assert(smoothers_->levels()==lhs_->levels() || matrices_->levels()==matrices_->maxlevels()); |
| 861 |
assert(smoothers_->levels()+1==lhs_->levels() || matrices_->levels()<matrices_->maxlevels()); |
| 862 |
|
| 863 |
if(smoother!=coarsest) |
| 864 |
for(++smoother, ++lhs, ++rhs; smoother != coarsest; ++smoother, ++lhs, ++rhs) |
| 865 |
smoother->pre(*lhs,*rhs); |
| 866 |
smoother->pre(*lhs,*rhs); |
| 867 |
} |
| 868 |
|
| 869 |
|
| 870 |
// The preconditioner might change x and b. So we have to |
| 871 |
// copy the changes to the original vectors. |
| 872 |
x = *lhs_->finest(); |
| 873 |
b = *rhs_->finest(); |
| 874 |
|
| 875 |
} |
| 876 |
template<class M, class X, class S, class PI, class A> |
| 877 |
std::size_t AMG<M,X,S,PI,A>::levels() |
| 878 |
{ |
| 879 |
return matrices_->levels(); |
| 880 |
} |
| 881 |
template<class M, class X, class S, class PI, class A> |
| 882 |
std::size_t AMG<M,X,S,PI,A>::maxlevels() |
| 883 |
{ |
| 884 |
return matrices_->maxlevels(); |
| 885 |
} |
| 886 |
|
| 887 |
/** \copydoc Preconditioner::apply */ |
| 888 |
template<class M, class X, class S, class PI, class A> |
| 889 |
void AMG<M,X,S,PI,A>::apply(Domain& v, const Range& d) |
| 890 |
{ |
| 891 |
LevelContext levelContext; |
| 892 |
|
| 893 |
if(additive) { |
| 894 |
*(rhs_->finest())=d; |
| 895 |
additiveMgc(); |
| 896 |
v=*lhs_->finest(); |
| 897 |
}else{ |
| 898 |
// Init all iterators for the current level |
| 899 |
initIteratorsWithFineLevel(levelContext); |
| 900 |
|
| 901 |
|
| 902 |
*levelContext.lhs = v; |
| 903 |
*levelContext.rhs = d; |
| 904 |
*levelContext.update=0; |
| 905 |
levelContext.level=0; |
| 906 |
|
| 907 |
mgc(levelContext); |
| 908 |
|
| 909 |
if(postSteps_==0||matrices_->maxlevels()==1) |
| 910 |
levelContext.pinfo->copyOwnerToAll(*levelContext.update, *levelContext.update); |
| 911 |
|
| 912 |
v=*levelContext.update; |
| 913 |
} |
| 914 |
|
| 915 |
} |
| 916 |
|
| 917 |
template<class M, class X, class S, class PI, class A> |
| 918 |
void AMG<M,X,S,PI,A>::initIteratorsWithFineLevel(LevelContext& levelContext) |
| 919 |
{ |
| 920 |
levelContext.smoother = smoothers_->finest(); |
| 921 |
levelContext.matrix = matrices_->matrices().finest(); |
| 922 |
levelContext.pinfo = matrices_->parallelInformation().finest(); |
| 923 |
levelContext.redist = |
| 924 |
matrices_->redistributeInformation().begin(); |
| 925 |
levelContext.aggregates = matrices_->aggregatesMaps().begin(); |
| 926 |
levelContext.lhs = lhs_->finest(); |
| 927 |
levelContext.update = update_->finest(); |
| 928 |
levelContext.rhs = rhs_->finest(); |
| 929 |
} |
| 930 |
|
| 931 |
template<class M, class X, class S, class PI, class A> |
| 932 |
bool AMG<M,X,S,PI,A> |
| 933 |
::moveToCoarseLevel(LevelContext& levelContext) |
| 934 |
{ |
| 935 |
|
| 936 |
bool processNextLevel=true; |
| 937 |
|
| 938 |
if(levelContext.redist->isSetup()) { |
| 939 |
levelContext.redist->redistribute(static_cast<const Range&>(*levelContext.rhs), |
| 940 |
levelContext.rhs.getRedistributed()); |
| 941 |
processNextLevel = levelContext.rhs.getRedistributed().size()>0; |
| 942 |
if(processNextLevel) { |
| 943 |
//restrict defect to coarse level right hand side. |
| 944 |
typename Hierarchy<Range,A>::Iterator fineRhs = levelContext.rhs++; |
| 945 |
++levelContext.pinfo; |
| 946 |
Transfer<typename OperatorHierarchy::AggregatesMap::AggregateDescriptor,Range,ParallelInformation> |
| 947 |
::restrictVector(*(*levelContext.aggregates), *levelContext.rhs, |
| 948 |
static_cast<const Range&>(fineRhs.getRedistributed()), |
| 949 |
*levelContext.pinfo); |
| 950 |
} |
| 951 |
}else{ |
| 952 |
//restrict defect to coarse level right hand side. |
| 953 |
typename Hierarchy<Range,A>::Iterator fineRhs = levelContext.rhs++; |
| 954 |
++levelContext.pinfo; |
| 955 |
Transfer<typename OperatorHierarchy::AggregatesMap::AggregateDescriptor,Range,ParallelInformation> |
| 956 |
::restrictVector(*(*levelContext.aggregates), |
| 957 |
*levelContext.rhs, static_cast<const Range&>(*fineRhs), |
| 958 |
*levelContext.pinfo); |
| 959 |
} |
| 960 |
|
| 961 |
if(processNextLevel) { |
| 962 |
// prepare coarse system |
| 963 |
++levelContext.lhs; |
| 964 |
++levelContext.update; |
| 965 |
++levelContext.matrix; |
| 966 |
++levelContext.level; |
| 967 |
++levelContext.redist; |
| 968 |
|
| 969 |
if(levelContext.matrix != matrices_->matrices().coarsest() || matrices_->levels()<matrices_->maxlevels()) { |
| 970 |
// next level is not the globally coarsest one |
| 971 |
++levelContext.smoother; |
| 972 |
++levelContext.aggregates; |
| 973 |
} |
| 974 |
// prepare the update on the next level |
| 975 |
*levelContext.update=0; |
| 976 |
} |
| 977 |
return processNextLevel; |
| 978 |
} |
| 979 |
|
| 980 |
template<class M, class X, class S, class PI, class A> |
| 981 |
void AMG<M,X,S,PI,A> |
| 982 |
::moveToFineLevel(LevelContext& levelContext, bool processNextLevel) |
| 983 |
{ |
| 984 |
if(processNextLevel) { |
| 985 |
if(levelContext.matrix != matrices_->matrices().coarsest() || matrices_->levels()<matrices_->maxlevels()) { |
| 986 |
// previous level is not the globally coarsest one |
| 987 |
--levelContext.smoother; |
| 988 |
--levelContext.aggregates; |
| 989 |
} |
| 990 |
--levelContext.redist; |
| 991 |
--levelContext.level; |
| 992 |
//prolongate and add the correction (update is in coarse left hand side) |
| 993 |
--levelContext.matrix; |
| 994 |
|
| 995 |
//typename Hierarchy<Domain,A>::Iterator coarseLhs = lhs--; |
| 996 |
--levelContext.lhs; |
| 997 |
--levelContext.pinfo; |
| 998 |
} |
| 999 |
if(levelContext.redist->isSetup()) { |
| 1000 |
// Need to redistribute during prolongateVector |
| 1001 |
levelContext.lhs.getRedistributed()=0; |
| 1002 |
Transfer<typename OperatorHierarchy::AggregatesMap::AggregateDescriptor,Range,ParallelInformation> |
| 1003 |
::prolongateVector(*(*levelContext.aggregates), *levelContext.update, *levelContext.lhs, |
| 1004 |
levelContext.lhs.getRedistributed(), |
| 1005 |
matrices_->getProlongationDampingFactor(), |
| 1006 |
*levelContext.pinfo, *levelContext.redist); |
| 1007 |
}else{ |
| 1008 |
*levelContext.lhs=0; |
| 1009 |
Transfer<typename OperatorHierarchy::AggregatesMap::AggregateDescriptor,Range,ParallelInformation> |
| 1010 |
::prolongateVector(*(*levelContext.aggregates), *levelContext.update, *levelContext.lhs, |
| 1011 |
matrices_->getProlongationDampingFactor(), |
| 1012 |
*levelContext.pinfo); |
| 1013 |
} |
| 1014 |
|
| 1015 |
|
| 1016 |
if(processNextLevel) { |
| 1017 |
--levelContext.update; |
| 1018 |
--levelContext.rhs; |
| 1019 |
} |
| 1020 |
|
| 1021 |
*levelContext.update += *levelContext.lhs; |
| 1022 |
} |
| 1023 |
|
| 1024 |
template<class M, class X, class S, class PI, class A> |
| 1025 |
bool AMG<M,X,S,PI,A>::usesDirectCoarseLevelSolver() const |
| 1026 |
{ |
| 1027 |
return IsDirectSolver< CoarseSolver>::value; |
| 1028 |
} |
| 1029 |
|
| 1030 |
template<class M, class X, class S, class PI, class A> |
| 1031 |
void AMG<M,X,S,PI,A>::mgc(LevelContext& levelContext){ |
| 1032 |
if(levelContext.matrix == matrices_->matrices().coarsest() && levels()==maxlevels()) { |
| 1033 |
// Solve directly |
| 1034 |
InverseOperatorResult res; |
| 1035 |
res.converged=true; // If we do not compute this flag will not get updated |
| 1036 |
if(levelContext.redist->isSetup()) { |
| 1037 |
levelContext.redist->redistribute(*levelContext.rhs, levelContext.rhs.getRedistributed()); |
| 1038 |
if(levelContext.rhs.getRedistributed().size()>0) { |
| 1039 |
// We are still participating in the computation |
| 1040 |
levelContext.pinfo.getRedistributed().copyOwnerToAll(levelContext.rhs.getRedistributed(), |
| 1041 |
levelContext.rhs.getRedistributed()); |
| 1042 |
solver_->apply(levelContext.update.getRedistributed(), |
| 1043 |
levelContext.rhs.getRedistributed(), res); |
| 1044 |
} |
| 1045 |
levelContext.redist->redistributeBackward(*levelContext.update, levelContext.update.getRedistributed()); |
| 1046 |
levelContext.pinfo->copyOwnerToAll(*levelContext.update, *levelContext.update); |
| 1047 |
}else{ |
| 1048 |
levelContext.pinfo->copyOwnerToAll(*levelContext.rhs, *levelContext.rhs); |
| 1049 |
solver_->apply(*levelContext.update, *levelContext.rhs, res); |
| 1050 |
} |
| 1051 |
|
| 1052 |
if (!res.converged) |
| 1053 |
coarsesolverconverged = false; |
| 1054 |
}else{ |
| 1055 |
// presmoothing |
| 1056 |
presmooth(levelContext, preSteps_); |
| 1057 |
|
| 1058 |
#ifndef DUNE_AMG_NO_COARSEGRIDCORRECTION |
| 1059 |
bool processNextLevel = moveToCoarseLevel(levelContext); |
| 1060 |
|
| 1061 |
if(processNextLevel) { |
| 1062 |
// next level |
| 1063 |
for(std::size_t i=0; i<gamma_; i++){ |
| 1064 |
mgc(levelContext); |
| 1065 |
if (levelContext.matrix == matrices_->matrices().coarsest() && levels()==maxlevels()) |
| 1066 |
break; |
| 1067 |
if(i+1 < gamma_){ |
| 1068 |
levelContext.matrix->applyscaleadd(-1., *levelContext.lhs, *levelContext.rhs); |
| 1069 |
} |
| 1070 |
} |
| 1071 |
} |
| 1072 |
|
| 1073 |
moveToFineLevel(levelContext, processNextLevel); |
| 1074 |
#else |
| 1075 |
*lhs=0; |
| 1076 |
#endif |
| 1077 |
|
| 1078 |
if(levelContext.matrix == matrices_->matrices().finest()) { |
| 1079 |
coarsesolverconverged = matrices_->parallelInformation().finest()->communicator().prod(coarsesolverconverged); |
| 1080 |
if(!coarsesolverconverged) |
| 1081 |
DUNE_THROW(MathError, "Coarse solver did not converge"); |
| 1082 |
} |
| 1083 |
// postsmoothing |
| 1084 |
postsmooth(levelContext, postSteps_); |
| 1085 |
|
| 1086 |
} |
| 1087 |
} |
| 1088 |
|
| 1089 |
template<class M, class X, class S, class PI, class A> |
| 1090 |
void AMG<M,X,S,PI,A>::additiveMgc(){ |
| 1091 |
|
| 1092 |
// restrict residual to all levels |
| 1093 |
typename ParallelInformationHierarchy::Iterator pinfo=matrices_->parallelInformation().finest(); |
| 1094 |
typename Hierarchy<Range,A>::Iterator rhs=rhs_->finest(); |
| 1095 |
typename Hierarchy<Domain,A>::Iterator lhs = lhs_->finest(); |
| 1096 |
typename OperatorHierarchy::AggregatesMapList::const_iterator aggregates=matrices_->aggregatesMaps().begin(); |
| 1097 |
|
| 1098 |
for(typename Hierarchy<Range,A>::Iterator fineRhs=rhs++; fineRhs != rhs_->coarsest(); fineRhs=rhs++, ++aggregates) { |
| 1099 |
++pinfo; |
| 1100 |
Transfer<typename OperatorHierarchy::AggregatesMap::AggregateDescriptor,Range,ParallelInformation> |
| 1101 |
::restrictVector(*(*aggregates), *rhs, static_cast<const Range&>(*fineRhs), *pinfo); |
| 1102 |
} |
| 1103 |
|
| 1104 |
// pinfo is invalid, set to coarsest level |
| 1105 |
//pinfo = matrices_->parallelInformation().coarsest |
| 1106 |
// calculate correction for all levels |
| 1107 |
lhs = lhs_->finest(); |
| 1108 |
typename Hierarchy<Smoother,A>::Iterator smoother = smoothers_->finest(); |
| 1109 |
|
| 1110 |
for(rhs=rhs_->finest(); rhs != rhs_->coarsest(); ++lhs, ++rhs, ++smoother) { |
| 1111 |
// presmoothing |
| 1112 |
*lhs=0; |
| 1113 |
smoother->apply(*lhs, *rhs); |
| 1114 |
} |
| 1115 |
|
| 1116 |
// Coarse level solve |
| 1117 |
#ifndef DUNE_AMG_NO_COARSEGRIDCORRECTION |
| 1118 |
InverseOperatorResult res; |
| 1119 |
pinfo->copyOwnerToAll(*rhs, *rhs); |
| 1120 |
solver_->apply(*lhs, *rhs, res); |
| 1121 |
|
| 1122 |
if(!res.converged) |
| 1123 |
DUNE_THROW(MathError, "Coarse solver did not converge"); |
| 1124 |
#else |
| 1125 |
*lhs=0; |
| 1126 |
#endif |
| 1127 |
// Prologate and add up corrections from all levels |
| 1128 |
--pinfo; |
| 1129 |
--aggregates; |
| 1130 |
|
| 1131 |
for(typename Hierarchy<Domain,A>::Iterator coarseLhs = lhs--; coarseLhs != lhs_->finest(); coarseLhs = lhs--, --aggregates, --pinfo) { |
| 1132 |
Transfer<typename OperatorHierarchy::AggregatesMap::AggregateDescriptor,Range,ParallelInformation> |
| 1133 |
::prolongateVector(*(*aggregates), *coarseLhs, *lhs, 1.0, *pinfo); |
| 1134 |
} |
| 1135 |
} |
| 1136 |
|
| 1137 |
|
| 1138 |
/** \copydoc Preconditioner::post */ |
| 1139 |
template<class M, class X, class S, class PI, class A> |
| 1140 |
void AMG<M,X,S,PI,A>::post([[maybe_unused]] Domain& x) |
| 1141 |
{ |
| 1142 |
// Postprocess all smoothers |
| 1143 |
typedef typename Hierarchy<Smoother,A>::Iterator Iterator; |
| 1144 |
typedef typename Hierarchy<Domain,A>::Iterator DIterator; |
| 1145 |
Iterator coarsest = smoothers_->coarsest(); |
| 1146 |
Iterator smoother = smoothers_->finest(); |
| 1147 |
DIterator lhs = lhs_->finest(); |
| 1148 |
if(smoothers_->levels()>0) { |
| 1149 |
if(smoother != coarsest || matrices_->levels()<matrices_->maxlevels()) |
| 1150 |
smoother->post(*lhs); |
| 1151 |
if(smoother!=coarsest) |
| 1152 |
for(++smoother, ++lhs; smoother != coarsest; ++smoother, ++lhs) |
| 1153 |
smoother->post(*lhs); |
| 1154 |
smoother->post(*lhs); |
| 1155 |
} |
| 1156 |
lhs_ = nullptr; |
| 1157 |
update_ = nullptr; |
| 1158 |
rhs_ = nullptr; |
| 1159 |
} |
| 1160 |
|
| 1161 |
template<class M, class X, class S, class PI, class A> |
| 1162 |
template<class A1> |
| 1163 |
void AMG<M,X,S,PI,A>::getCoarsestAggregateNumbers(std::vector<std::size_t,A1>& cont) |
| 1164 |
{ |
| 1165 |
matrices_->getCoarsestAggregatesOnFinest(cont); |
| 1166 |
} |
| 1167 |
|
| 1168 |
} // end namespace Amg |
| 1169 |
|
| 1170 |
struct AMGCreator{ |
| 1171 |
template<class> struct isValidMatrix : std::false_type{}; |
| 1172 |
template<class T, int n, int m, class A> struct isValidMatrix<BCRSMatrix<FieldMatrix<T,n,m>, A>> : std::true_type{}; |
| 1173 |
|
| 1174 |
template<class OP> |
| 1175 |
std::shared_ptr<Dune::Preconditioner<typename OP::element_type::domain_type, typename OP::element_type::range_type> > |
| 1176 |
makeAMG(const OP& op, const std::string& smoother, const Dune::ParameterTree& config) const |
| 1177 |
{ |
| 1178 |
DUNE_THROW(Dune::Exception, "Operator type not supported by AMG"); |
| 1179 |
} |
| 1180 |
|
| 1181 |
template<class M, class X, class Y> |
| 1182 |
std::shared_ptr<Dune::Preconditioner<X,Y> > |
| 1183 |
makeAMG(const std::shared_ptr<MatrixAdapter<M,X,Y>>& op, const std::string& smoother, |
| 1184 |
const Dune::ParameterTree& config) const |
| 1185 |
{ |
| 1186 |
using OP = MatrixAdapter<M,X,Y>; |
| 1187 |
|
| 1188 |
if(smoother == "ssor") |
| 1189 |
return std::make_shared<Amg::AMG<OP, X, SeqSSOR<M,X,Y>>>(op, config); |
| 1190 |
if(smoother == "sor") |
| 1191 |
return std::make_shared<Amg::AMG<OP, X, SeqSOR<M,X,Y>>>(op, config); |
| 1192 |
if(smoother == "jac") |
| 1193 |
return std::make_shared<Amg::AMG<OP, X, SeqJac<M,X,Y>>>(op, config); |
| 1194 |
if(smoother == "gs") |
| 1195 |
return std::make_shared<Amg::AMG<OP, X, SeqGS<M,X,Y>>>(op, config); |
| 1196 |
if(smoother == "ilu") |
| 1197 |
return std::make_shared<Amg::AMG<OP, X, SeqILU<M,X,Y>>>(op, config); |
| 1198 |
else |
| 1199 |
DUNE_THROW(Dune::Exception, "Unknown smoother for AMG"); |
| 1200 |
} |
| 1201 |
|
| 1202 |
template<class M, class X, class Y, class C> |
| 1203 |
std::shared_ptr<Dune::Preconditioner<X,Y> > |
| 1204 |
makeAMG(const std::shared_ptr<OverlappingSchwarzOperator<M,X,Y,C>>& op, const std::string& smoother, |
| 1205 |
const Dune::ParameterTree& config) const |
| 1206 |
{ |
| 1207 |
using OP = OverlappingSchwarzOperator<M,X,Y,C>; |
| 1208 |
|
| 1209 |
auto cop = std::static_pointer_cast<const OP>(op); |
| 1210 |
|
| 1211 |
if(smoother == "ssor") |
| 1212 |
return std::make_shared<Amg::AMG<OP, X, BlockPreconditioner<X,Y,C,SeqSSOR<M,X,Y>>,C>>(cop, config, op->getCommunication()); |
| 1213 |
if(smoother == "sor") |
| 1214 |
return std::make_shared<Amg::AMG<OP, X, BlockPreconditioner<X,Y,C,SeqSOR<M,X,Y>>,C>>(cop, config, op->getCommunication()); |
| 1215 |
if(smoother == "jac") |
| 1216 |
return std::make_shared<Amg::AMG<OP, X, BlockPreconditioner<X,Y,C,SeqJac<M,X,Y>>,C>>(cop, config, op->getCommunication()); |
| 1217 |
if(smoother == "gs") |
| 1218 |
return std::make_shared<Amg::AMG<OP, X, BlockPreconditioner<X,Y,C,SeqGS<M,X,Y>>,C>>(cop, config, op->getCommunication()); |
| 1219 |
if(smoother == "ilu") |
| 1220 |
return std::make_shared<Amg::AMG<OP, X, BlockPreconditioner<X,Y,C,SeqILU<M,X,Y>>,C>>(cop, config, op->getCommunication()); |
| 1221 |
else |
| 1222 |
DUNE_THROW(Dune::Exception, "Unknown smoother for AMG"); |
| 1223 |
} |
| 1224 |
|
| 1225 |
template<class M, class X, class Y, class C> |
| 1226 |
std::shared_ptr<Dune::Preconditioner<X,Y> > |
| 1227 |
makeAMG(const std::shared_ptr<NonoverlappingSchwarzOperator<M,X,Y,C>>& op, const std::string& smoother, |
| 1228 |
const Dune::ParameterTree& config) const |
| 1229 |
{ |
| 1230 |
using OP = NonoverlappingSchwarzOperator<M,X,Y,C>; |
| 1231 |
|
| 1232 |
if(smoother == "ssor") |
| 1233 |
return std::make_shared<Amg::AMG<OP, X, NonoverlappingBlockPreconditioner<C,SeqSSOR<M,X,Y>>,C>>(op, config, op->getCommunication()); |
| 1234 |
if(smoother == "sor") |
| 1235 |
return std::make_shared<Amg::AMG<OP, X, NonoverlappingBlockPreconditioner<C,SeqSOR<M,X,Y>>,C>>(op, config, op->getCommunication()); |
| 1236 |
if(smoother == "jac") |
| 1237 |
return std::make_shared<Amg::AMG<OP, X, NonoverlappingBlockPreconditioner<C,SeqJac<M,X,Y>>,C>>(op, config, op->getCommunication()); |
| 1238 |
if(smoother == "gs") |
| 1239 |
return std::make_shared<Amg::AMG<OP, X, NonoverlappingBlockPreconditioner<C,SeqGS<M,X,Y>>,C>>(op, config, op->getCommunication()); |
| 1240 |
if(smoother == "ilu") |
| 1241 |
return std::make_shared<Amg::AMG<OP, X, NonoverlappingBlockPreconditioner<C,SeqILU<M,X,Y>>,C>>(op, config, op->getCommunication()); |
| 1242 |
else |
| 1243 |
DUNE_THROW(Dune::Exception, "Unknown smoother for AMG"); |
| 1244 |
} |
| 1245 |
|
| 1246 |
template<typename OpTraits, typename OP> |
| 1247 |
std::shared_ptr<Dune::Preconditioner<typename OpTraits::domain_type, |
| 1248 |
typename OpTraits::range_type>> |
| 1249 |
operator() (OpTraits opTraits, const std::shared_ptr<OP>& op, const Dune::ParameterTree& config, |
| 1250 |
std::enable_if_t<isValidMatrix<typename OpTraits::matrix_type>::value,int> = 0) const |
| 1251 |
{ |
| 1252 |
using field_type = typename OpTraits::matrix_type::field_type; |
| 1253 |
using real_type = typename FieldTraits<field_type>::real_type; |
| 1254 |
if (!std::is_convertible<field_type, real_type>()) |
| 1255 |
DUNE_THROW(UnsupportedType, "AMG needs field_type(" << |
| 1256 |
className<field_type>() << |
| 1257 |
") to be convertible to its real_type (" << |
| 1258 |
className<real_type>() << |
| 1259 |
")."); |
| 1260 |
std::string smoother = config.get("smoother", "ssor"); |
| 1261 |
// we can irgnore the OpTraits here. As the AMG can only work |
| 1262 |
// with actual matrices, the operator op must be of type |
| 1263 |
// MatrixAdapter or *SchwarzOperator. In any case these |
| 1264 |
// operators provide all necessary information about matrix, |
| 1265 |
// domain and range type |
| 1266 |
return makeAMG(op, smoother, config); |
| 1267 |
} |
| 1268 |
|
| 1269 |
template<typename OpTraits, typename OP> |
| 1270 |
std::shared_ptr<Dune::Preconditioner<typename OpTraits::domain_type, |
| 1271 |
typename OpTraits::range_type>> |
| 1272 |
operator() (OpTraits opTraits, const std::shared_ptr<OP>& op, const Dune::ParameterTree& config, |
| 1273 |
std::enable_if_t<!isValidMatrix<typename OpTraits::matrix_type>::value,int> = 0) const |
| 1274 |
{ |
| 1275 |
DUNE_THROW(UnsupportedType, "AMG needs a FieldMatrix as Matrix block_type"); |
| 1276 |
} |
| 1277 |
}; |
| 1278 |
|
| 1279 |
DUNE_REGISTER_PRECONDITIONER("amg", AMGCreator()); |
| 1280 |
} // end namespace Dune |
| 1281 |
|
| 1282 |
#endif |